Synergy of unsupervised and supervised machine learning methods for the segmentation of the graphite particles in the microstructure of ductile iron

Abstract

In this research, a machine learning (ML) approach was used for automating the measurement of the area fraction of the graphite particles in the microstructure of ductile iron. The machine learning pipeline includes an unsupervised learning step followed by a supervised learning step. The unsupervised learning step uses K-means clustering to produce the segmentation masks, which are used as the training data for the supervised learning step. For the supervised learning step, the U-Net architecture, which is a fully convolutional neural network, is used to build the model and to predict the area fraction of graphite in the microstructure of ductile iron. To prevent overfitting, data augmentation is used to generate more samples for the dataset. The focal loss function is applied to the training process to reduce the effect of the unbalance data in the training dataset and to enhance the classification accuracy of the model. The developed ML methods show promising results in terms of the accuracy of the graphite segmentation and the matching with the measured values by human experts. This outcome can be readily tailored for the microstructural analysis of other composite materials, which can lead to better design of materials with enhanced properties.